Improving Telescope Operations via Explainable AI and Reinforcement Learning

Light travels with a finite speed. As it takes eight minutes for the light from the Sun to reach us, the light that we see on Earth is actually an image of the Sun eight minutes ago. By extending this logic, modern-day astronomy aims to observe objects that are increasingly far away from us. Since the light from these astronomical objects can take billions of years to reach us. By studying distant objects, we can essentially go back in time and study the Universe when it was much younger However, studying distant objects in the Universe also comes with an actual cost. Distant objects are faint, and we need large telescopes to collect sufficient photons from them. However, the operation of such cutting-edge large telescopes can be very costly. For example, a state-of-the-art telescope with 8m in aperture costs AUD 80K per night to operate (or $1.5 per second!). And the coming 30m telescopes will cost even an order of magnitude more!

Yet, the current operation of telescopes leaves a lot to be desired. First, the operation of operations often still relies on experienced human operators. On top of that, the sequence with which astronomical targets are observed is often predefined by astronomers. While this model has worked well, it is clearly not optimal. The ultimate telescope operation and scheduling should, in principle, take into account the real-time information, including the weather, the meta-data from various telescope detectors, and update the scheduling as the observation is being carried out. Fortuitously, most of the telescope metadata from decades of real-life observations are all well documented; this digital record can now serve as a fantastic training set and sandbox to explore modern machine learning applications. This project will explore how explainable AI can leverage historical records to assist human operators. We will also explore how future astronomical survey scheduling can be improved via reinforcement learning. 

 

Goals

Develop explainable AI that can take into account real-time telescope meta-data to assist human operators. Explore telescope scheduling as a reinforcement or bandit problem.

 

Background Literature

https://ui.adsabs.harvard.edu/abs/2019AJ....157..151N

https://ui.adsabs.harvard.edu/abs/2020arXiv201103132G

https://ui.adsabs.harvard.edu/abs/2021arXiv210700048G

 

Requirements

Python programming (Pytorch, Tensorflow) and experience in machine learning and data science are essential.  Familiarity with explainable-AI and reinforcement learning is desirable.